Parameters Reference

Detailed explanation of generation parameters for LTX-2 / 2.3, LTX-2.5, and MiniMax H3.

Model
Resolution
1. Width & Height
2. Common Aspect Ratios
Frames
1. Number of Frames
2. FPS (Frames Per Second)
Quality Settings
1. Inference Steps
2. Guidance Scale
Seed
1. Random Seed
Image-to-Video Settings
1. Image Strength
Prompts
1. Main Prompt
2. Negative Prompt
Presets
Audio Settings
Performance Notes
1. Memory Usage
2. Generation Time

Model

Select the model in Preferences. LTX-2 / 2.3 notapalindrome packs use mlx-video-with-audio. LTX-2.3 12GB and LTX-2.5 use ltx-2-mlx. MiniMax H3 uses the native h3 binary.

Model	Parameters	Size	Vocoder	Notes
LTX-2 Unified (`notapalindrome/ltx2-mlx-av`)	19B	~42GB	Standard	Original unified model
LTX-2.3 Unified Beta (`notapalindrome/ltx23-mlx-av`)	19B	~48GB	BigVGAN	Distilled beta model
LTX-2.3 Distilled Q4 Beta (`notapalindrome/ltx23-mlx-av-q4`)	19B (Q4)	~22GB	BigVGAN	Quantized; default for new installs except ≤16GB Macs
LTX-2.3 12GB RAM Optimized (`baa-ai/LTX-2.3-22B-RAM-12GB-MLX`)	22B mixed	~19GB	BigVGAN	`ltx-2-mlx`; Gemma 3 via `--gemma`; default on ≤16GB; vendor ~14GB claim
LTX-2.5 Distilled (`notapalindrome/ltx25-mlx`)	22B	~110GB	BigVGAN	Gemma 4 in `gemma4-12b-ltx-v1/`; LoRA bundled; `ltx-2-mlx` 0.15+; 8 steps
LTX-2.5 Dev two-stage (`ltx25_dev`)	22B	same pack	BigVGAN	`--two-stage` + `transformer-dev` + fuse bundled `ltx-2.5-22b-distilled-lora-450-bf16`; stage-1 30 / CFG 3 / stage-2 3
LTX-2.5 Distilled Q8 DiT	22B (Q8 DiT)	~100GB + ~21GB	BigVGAN	overlay `notapalindrome/ltx25-mlx-ditq8`
MiniMax H3 (`MiniMaxAI/MiniMax-H3`)	33B	~144GB	H3 audio VAE	Native `h3.c`; 24 fps; frames snap to 5+17n

LTX-2 / 2.3 notapalindrome packs use a 2-stage pipeline (half-res, then refine) with synchronized stereo audio. ltx-2-mlx distilled (2.5 Distilled / Q8 DiT and the 12GB pack) is a fixed 8-step pass (CFG=1). ltx25_dev is the official two-stage path (dev DiT + CFG, then distilled-LoRA refine). H3 is 24 fps with frames snapped to 5+17n.

See Architecture for full technical details on how each model is loaded and decoded.

Resolution

Width & Height

Controls the dimensions of the generated video in pixels.

Setting	Range	Default
Width	320-1024	512
Height	320-768	320

Tips:

Both dimensions should be divisible by 64 for best results
Higher resolutions require more memory and time
Start with smaller sizes (512x320) for testing

Common Aspect Ratios

Aspect	Width × Height	Use Case
16:10	512×320, 640×400, 768×480	Standard
16:9	768×432, 1024×576	Widescreen
9:16	384×640, 432×768	Portrait/Mobile
1:1	512×512	Social media
2.4:1	768×320	Cinematic

Frames

Number of Frames

Total frames to generate. More frames = longer video.

Setting	Range	Default
Frames	17-257	49

Frames are automatically adjusted to be 8n+1 (9, 17, 25, 33, 41, 49… 121, etc.) as required by the model.

Duration calculation:

Duration (seconds) = Frames ÷ FPS

Examples at 24 FPS:

49 frames = ~2 seconds
97 frames = ~4 seconds
121 frames = ~5 seconds

FPS (Frames Per Second)

Playback speed of the generated video.

Setting	Range	Default
FPS	12-30	24

Tips:

24 FPS: Cinematic feel
30 FPS: Smooth motion
12 FPS: Animation style
For synchronized speech/lip-sync, use 24 FPS

Quality Settings

Inference Steps

Number of denoising steps per stage.

Setting	Range	Default
Steps	8-50	28

LTX-2 / 2.3 notapalindrome catalog models use Lightricks’ distilled sigma schedule: 8 stage-1 steps + 3 stage-2 steps (11 total). ltx-2-mlx distilled (2.5 Distilled / Q8 DiT and ltx23_12gb) is a fixed 8-step schedule (CFG=1). ltx25_dev uses the slider as stage-1 steps (default 30, recommended 20–40) and CFG (default 3); stage-2 is 3. H3 uses Fast / Default / Reference presets (4/50/1, 20/45/2, 50/50/1). The inference-steps slider is ignored for the LTX distilled catalogs.

Guidance Scale

How closely the model follows your prompt (CFG - Classifier-Free Guidance).

Setting	Range	Default
Guidance	1.0-10.0	4.0

Effects:

1.0: Minimal guidance
3-5: Balanced adherence (recommended)
6+: Stronger prompt following

Recommended: 3.5 - 5.0 for LTX-2

Seed

Random Seed

Controls the randomness of generation.

Setting	Range	Default
Seed	0-2147483647	Random

Usage:

Same seed + same parameters = identical output
Leave empty for random seed each time
Save the seed of good results to reproduce them
Try nearby seeds for variations

Image-to-Video Settings

When using image-to-video mode:

Image Strength

How much the source image influences the output.

Setting	Range	Default
Strength	0.0-1.0	1.0

1.0: Strong influence, first frame closely matches image
0.5: Moderate influence
0.0: Image is ignored (text-to-video mode)

Prompts

Main Prompt

Describe what you want to see in the video.

Best practices:

Be descriptive and specific
Include camera movement if desired
Describe motion and action
Mention lighting and atmosphere

Negative Prompt

Describe what you want to avoid.

Common negative prompts:

worst quality, low quality, blurry, 
jittery motion, distorted, 
watermark, text, logo,
static, still image

Presets

Presets provide quick access to common configurations:

Quick Preview

512×320, 49 frames, 20 steps, guidance 4.0

Fast testing, ~2-3 minutes

Standard

768×512, 97 frames, 28 steps, guidance 4.0

Good balance, ~5-7 minutes

High Quality

768×512, 121 frames, 28 steps, guidance 4.0

Best results, longer duration

Portrait

512×768, 97 frames, 28 steps, guidance 4.0

Vertical format for mobile

Square

512×512, 97 frames, 28 steps, guidance 4.0

Social media format

Audio Settings

Voiceover Source

Choose between local or cloud-based text-to-speech.

Source	Quality	Speed	Requirements
MLX Audio (Local)	Good	Fast	None (free)
ElevenLabs (Cloud)	Excellent	Fast	API key

Voice Options

MLX Audio Voices:

Heart, Bella, Nova, Sky (US Female)
Adam, Echo (US Male)
Alice, Emma (UK Female)
Daniel, George (UK Male)

ElevenLabs Voices:

Rachel, Domi, Bella, Elli (Female)
Antoni, Josh, Arnold, Adam, Sam (Male)

Background Music

Requires ElevenLabs API key. Music is generated to match video length automatically.

Setting	Value
Volume (music only)	30%
Volume (with voiceover)	20%
Format	AAC 192kbps

Music Genres

54 presets across 9 categories. Each genre has a carefully crafted prompt optimized for ElevenLabs Music API.

Categories:

Electronic (8 genres)
Hip-Hop/R&B (6 genres)
Rock (6 genres)
Pop (4 genres)
Jazz/Blues (5 genres)
Classical/Cinematic (6 genres)
World (5 genres)
Country/Folk (4 genres)
Functional/Mood (10 genres)

Performance Notes

Memory Usage

LTX-2 is a 19B parameter model requiring significant unified memory:

Resolution	Approximate Memory
512×320	~24-28GB
768×512	~28-32GB
1024×768	~35-40GB+

Generation Time

Rough estimates on Apple Silicon:

Mac	Quick Preview	Standard
M1 (16GB)	May fail	Not recommended
M1 Pro (32GB)	4-5 min	8-10 min
M2 Max (64GB)	2-3 min	5-7 min
M3 Max (128GB)	1-2 min	3-5 min

Times vary based on:

Chip generation and core count
Unified memory amount
Other running applications
Resolution and frame count

32GB RAM minimum required. Macs with 16GB will likely fail or be extremely slow. 64GB+ recommended for comfortable usage.